Definition
An AI system that helps a human understand or change software while the human remains the primary driver of the workflow. Typical assistance includes code completion, explanation, search, question answering, refactoring suggestions, test generation, and proposed edits.
The defining feature is the allocation of control. The human selects the immediate problem, decides what context to provide, evaluates the response, and advances the work. The assistant may be fast and capable, but it is participating inside a human-led sequence rather than independently managing a multi-step objective.
A moving product boundary
AI copilot became a common product metaphor for this arrangement, but vendor names are unreliable classifiers. Many products now combine inline assistance with background agent modes. The same product can act as an assistant in one session and as an agent in another. Classify the actual behavior: who chooses the next action, who runs tools, how long the system proceeds without intervention, and who can change external state.
Distinguish it from nearby terms
A code generator may produce a snippet in one response. A coding assistant supports a human-led workflow. A coding agent can choose and execute a sequence of repository and tool actions toward an outcome. These categories describe control patterns, not a permanent ranking of products.
Check your understanding
An editor offers autocomplete while you type, then lets you hand a ticket to a background process that opens a pull request. Which mode is the coding assistant? The autocomplete interaction is. The background, multi-step execution is better described as a coding agent, even though both came from the same product.